Background
Gastric cancer is a major health problem worldwide. It ranks second in all causes of death from cancer, with about 700,000 confirmed deaths annually [
1,
2]. In Japan, although its mortality ranks also second and has decreased in recent years, it still has the highest incidence despite advances in prevention and treatment [
3]. While the internationally accepted standard treatment for patients with potentially resectable disease was surgery alone [
4,
5], meta-analyses of adjuvant chemotherapy for gastric cancer during the last few decades have shown reductions in mortality up to 18% [
6,
7]. However, these reductions were considered insufficient to change clinical practice.
Recently, the effectiveness of specific regimens for resectable gastric and/or gastroesophageal cancer has been verified in large clinical trials. The chemoradiation therapy (INT-0116) in the US in 2001 [
8], the perioperative chemotherapy (MAGIC) in Europe in 2006 [
9], and the postoperative chemotherapy (ACTS-GC) in Japan in 2007 [
10,
11] improved significantly overall survival (OS), and relapse-free survival (RFS) or progression-free survival (PFS), compared to surgery alone.
These studies have led to a new phase in the treatment of gastric cancer, even though there are several issues under discussion concerning them [
5,
12,
13]. Postoperative chemoradiotherapy, perioperative triplet-chemotherapy, and postoperative S-1 mono-chemotherapy are now the standard therapies in the US, Europe and Japan, respectively [
5,
12]. Also, the status of adjuvant treatment of gastric cancer has been evolving to improve and optimize the current standard of care across national boundaries.
Under these circumstances, from a perspective of healthcare policy, in choosing the best treatment among the different options available, clinical benefits of treatments should be balanced against the effects on costs, since rapid growth in healthcare expenditures creates an unsustainable burden. However, economic evaluation of adjuvant therapy for gastric cancer has been greatly lacking.
Our objective was to estimate the cost-effectiveness of adjuvant S-1 therapy in Japan. This study would provide basic information on the cost-effectiveness of adjuvant therapy for gastric cancer in Japan.
Discussion
From the perspective of the National Health Insurance in Japan, this cost-effectiveness analysis showed that S-1 adjuvant therapy for gastric cancer gained LYs and QALYs, while it increased costs, compared with surgery alone (Table
2). The ICER of S-1 therapy can be ranked close to the top of the league table of cost-utility in oncology [
33]. There is some consensus about the threshold of willingness to pay for additional QALY internationally (e.g., $50,000 in the US, £30,000 in the UK, or AUS $42,000 in Australia) [
34]. A recent review suggested that the plausible threshold is $109,000/QALY, rather than $50,000/QALY [
35]. In Japan, the social value (i.e., willingness to pay) for QALY gained was estimated to be from $53,000 to $56,000 by a nationwide mail survey using conjoint analysis [
36]. Since the ICER of S-1 therapy is far below these thresholds, it is considered acceptable.
There has been little evidence on economic evaluation of adjuvant therapy for gastric cancer. A cost-effectiveness analysis evaluating postoperative chemoradiotherapy for gastric cancer in the US showed that the incremental cost-effectiveness ratio was $38,400 per QALY gained [
37]. This ratio is 14 times higher and less efficient than that in our study, although several factors such as clinical practice patterns and relative costs should be considered in transferring evaluation data [
14]. Moreover, since there is no genuine utility information in calculating QALY in the report [
37], its validity and plausibility would be questionable.
The results of this study are subject to uncertainty and assumptions. To estimate stochastic uncertainty of ICER due to sampling variation or error, probabilistic sensitivity analyses [
14,
31,
32] were performed (Table
2, Figure
2). Cost-effectiveness scatter plots showed that all points of ICERs were located under the diagonal line indicating $50,000/QALY. CEAC and NMB curves give more information. If a decision-maker was willing to pay $6,220 to achieve an additional QALY, the likelihood of S-1 therapy being acceptable as cost-effective was 95% (Figure
2B). The NMB curve shows that S-1 therapy was beneficial, if a decision-maker was willing to pay $2,782 (Figure
2C). These values are extremely low compared with the thresholds (e.g., $50,000).
The time horizon is an important issue to sufficiently capture relevant costs and health outcomes of S-1 adjuvant therapy. The observation period of the ACTS-GC, 5 years was limited. While most costs were incurred mainly in the observational period, LYs gained would continue after it. In this study, a simulation model was used to extrapolate its results. There is a variety of ways for simulation [
18], but no uniform methodology available. We used the Boag model, which is indicated to be predictive for prognosis of gastric cancer [
17]. In a sensitivity analysis, the ICER of the observational period was much higher than that of over lifetime (the base case), but it is very low compared with the thresholds. Also, the results of other simulation methods indicated similar results. The exclusion of end-of-life costs due to gastric cancer slightly increased the ICER, but it still remained far under the threshold (Table
4). These analyses show the robustness of this study.
The key drivers of cost-effectiveness results of S-1 are mainly the acquisition cost of S-1 and the costs related to recurrence and death. The S-1 therapy partly offset the acquisition cost of S-1 by the savings achieved by reduction of these costs. In one-way sensitivity analysis (Table
4), varying recurrence rates and costs of recurrence and end-of-lie did not have substantial impact on cost-effectiveness. Varying acquisition cost, which was the other cost driver, also did not have major impact on cost-effectiveness (Table
4). The sensitivity analysis of total cost corresponded with these results.
Cost-effectiveness analysis using QALYs offers the opportunity to consider both quantity and quality of survival. However, no substantial difference in ICERs was observed between cost per LY gained and QALY gained (Table
2). In this study, utility values were derived from a relatively small number of patients with gastric cancer, but this is the first study which directly evaluated the utilities among patients with gastric cancer. These values are similar to those observed for general cancer (i.e, 0.89 after surgery and 0.44 for metastasis) in the Canadian survey among the general population [
38]. The sensitivity analysis on range of utility values for remission after surgery and metastasis revealed no major change in cost-effectiveness (Table
4). In a sizable fraction of cost-effectiveness analyses, utility weighting was indicated not to substantially alter the estimated cost-effectiveness of an intervention [
39]. It is thus suggested that sensitivity analyses using ad hoc adjustment or weight from the literature may be sufficient. Our results support this conclusion.
The impact of discounting for the time value of money on the results was examined extensively by two-way sensitivity analysis. Although ICERs were more sensitive to effectiveness discounting than cost discounting, there was no substantial change in cost-effectiveness. The main reason is likely to be that major costs were incurred during the early phase of follow-up and improved survival continued for a relatively long time.
There are additional limitations in the analysis that should be commented on. First, the perspective of this analysis is that of a payer for healthcare, rather than a society. From a societal perspective, the range of costs is broader and includes other costs such as indirect costs. Since S-1 therapy increased OS and decreased recurrence, these factors would reduce indirect costs and decrease its ICER.
Second, the issue of generalizability of this study to other countries should be carefully examined. S-1 is widely used in Asian countries (e.g., Japan, Korea, Singapore and China). However, it is difficult to determine the relative effectiveness of S-1, compared with the preoperative chemoradiotherapy in the US and the preoperative triplet-chemotherapy in Europe, since there is no direct comparison among them [
8‐
10]. Moreover, there are several critical arguments around these studies. For example, the INT-0116 study attracted some criticism on the grounds of poor standardization of surgery and insufficient extended dissection of regional lymph nodes [
5]. Thus it was argued that the chemoradiation component of the adjuvant treatment had compensated for less-than-ideal surgery. On the other hand, the quality of the MAGIC trial was pointed out to be much poorer than that of the INT-0116 study, in the areas of active quality control of surgery, data management, and compliance with protocol [
12]. As to S-1, a difference in S-1 phamacokinetics was observed between Asians and Caucasians [
13].
Recently, although the subjects did no have resectable gastric cancer like in this study, but advanced gastric cancer, the First-Line Advance Gastric Cancer Study (FLAGS) [
40], a multinational trial, showed that cisplatin/S-1 was statistically non-inferior in overall mortality to cisplatin/5-FU and showed a significantly improved safety profile in Western countries. While S-1 is now approved by the EMEA in European countries, an international head-to-head comparison between S-1 therapy and the Western standard therapies will be required to confirm relative effectiveness and cost-effectiveness of S-1 therapy.
Appendix: the method of QALY calculation
A.1 Calculation of QALY
QALYi (u), defined as the QALY at year i, was calculated by the following Equation (
1), in which uNR represents the utility value of no relapse and uR represents the utility value of relapse.
(1)
If d is the discount rate, the equation becomes QALY(u)=Σid(i-1) × QALYi(u).
The mean rate of survival was calculated as the area under the curve (AUC) of OS, and the mean rate of relapse-free survival was calculated as the AUC of RFS, using the trapezoidal approximation rule.
A.2 Estimate of survival curves of lifetime OS
When estimating the survival curves of lifetime OS, it was assumed that some patients in this study would be cured in response to treatment. This model is called the Boag (cure) model or mixture cure model. This statistical model assumes a mixed distribution of survival time among cured patients and uncured patients.
Y is defined as a variable indicating the presence or absence of cure in patients. Y = 0 stands for cure, and Y = 1 stands for non-cure. If p is defined as the probability of non-cure as represented by p = Pr(Y = 1), and T is a random variable indicating the survival time, the cumulative distribution function of T is represented by the following Equation (
2).
(2)
It was assumed that no events occur because of cure in cured patients. In other words, if Pr(T ≤ t|Y=0) = 0, the distribution function would be represented by Equation (
2). This is referred to as a cure model.
In the cure model, the probability density function f(t) and survival function S(t) are represented by the following Equations (
4).
(4)
A logistic regression model was assumed to calculate the probability of non-cure
p. In this model,
p is calculated by Equation (
5), in which z is a covariance vector, x = (1,z)
' (
' stands for vector transposition), and b is a regression coefficient vector of covariance.
(5)
The Boag model [
15] assumes a log-normal distribution for the survival time of uncured patients, but a log-logistic distribution was assumed in the present study. Furthermore, sensitivity analysis was also performed assuming a log-normal distribution and a Weibull distribution, and the maximum likelihood method was used to estimate the parameters using observational data of the ACTS-GC trial [
10,
11]. The goodness of fit of the model was evaluated with Akaike’s information criteria (AIC). A log-logistic distribution has two parameters
θ = (γ, λ)′, and the survivor function is as follows:
(6)
The statistical software package SAS (version 9.2) was used to fit the data to the aforementioned models, and the probabilities of non-cure (p) were estimated to be 0.306 and 0.422 in the S-1 group and surgery alone group, respectively. The log-logistic distribution parameters λ and γ were 0.9724 and 0.4121, respectively. The value of AIC for the log-logistic model was 1,678. Those for log-normal and Weibull models were 2,113 and 2,117, respectively. The programs used to estimate the model parameters were the SAS macro for survival models with a cured fraction (Mixture Cure Models).
To examine the validity of the log-logistic model, the distribution of survival time of cured patients was also analyzed using data on patients with gastric cancer obtained from the Cancer Institute Hospital (1946–2004), which has an open database [
19]. The approach used was as follows: First, data on patients who met the following 6 eligibility criteria corresponding to the ACTS-GC trial (n = 1,457) were extracted from all data (n = 13,740). The median age of the patients extracted from the database was 57 years, which was 6 years younger than the median age of 63 years in the ACTS-GC trial. Kaplan-Meier curves were plotted using the extracted patient data, defining only death from gastric cancer as an event. The curve reached a plateau after about 20 years (corresponding to an age of 77 years). These data were used for cure models assuming a Weibull distribution, log-normal distribution, and log-logistic distribution. The goodness of fit of the data as indicated by the AIC was best for the log-logistic distribution. While the value of AIC for the log-logistic model was 1,845, those for the log-normal and Weibull models were 2,071 and 2,105, respectively.
Eligibility criteria of the ACTS-GC trial
1)
A histologically confirmed diagnosis of gastric cancer
2)
Lymph-node dissection of D2 or greater, with a curability of A or B
3)
Stage II, IIIA, or IIIB disease
4)
No liver metastasis, hematogenous metastasis, or distant metastasis
6)
No previous treatment (chemotherapy, radiotherapy) received
Finally, the OS curve was constructed by combining the disease-specific survival curve (cure parametric model) and the disease-independent survival curve (the general population matched for age and sex of the subjects) based on the competing risk model. The actual calculation was done using a competitive risk model and the following Equation (
7), in which S
B(t) stands for the survival rate in the disease-specific survival curve (= cure model curve), S
C(t) stands for the survival rate of the general population in the disease-independent survival curve, and S
A(t) is the estimated rate of OS after the observation period. The structure of the OS curve was presented in Figure
1B.
Competing interest
MS reports receiving lectures fees from Taiho. All other authors: none to declare.
Authors’ contributions
AH: study concept and design, acquisition of economic data, analysis and interpretation of economic data, and preparation of manuscript.MS, SN: acquisition of subjects and/or clinical data, analysis and interpretation of clinical data. All authors read and approved the final manuscript.